- 1China University of Petroleum(East China), Qingdao, China
- 2State Key Laboratory of Deep Oil and Gas, China University of Petroleum (East China), Qingdao, China
- 3Southwest Measurement & Control Company, Sinopec Matrix Corporation, Sichuan, China
- 4Sinopec Matrix Corporation, Qingdao, China
- 5Oil and Gas Exploration Management Center, Sinopec Shengli Oilfield Company, Dongying, China
- 6School of Safety Engineering, China University of Mining and Technology, Xuzhou, China
Marine shales of X-block are featured with poor organic matter, over maturity and complex mineralogical assemblages, which collectively result in weak organic-related logging responses. As a consequence, conventional total organic carbon (TOC) evaluation methods exhibit substantial uncertainties associated with baseline calibration and parameter generalization, thereby limiting prediction accuracy and robustness.
To overcome these limitations, this study develops a physics-constrained deep learning framework for TOC prediction that integrates multi-source logging data using a hybrid DNN–LSTM–GPC architecture. High-resolution nuclear magnetic resonance (NMR) and electrical imaging logs are incorporated as primary data sources to extract multi-scale, organic-matter-sensitive features. These features are further integrated with conventional well logs to construct a comprehensive feature space that captures organic matter distribution, pore structure characteristics, and lithological variability. In addition, an improved ΔLogR model and region-specific rock-physics constraints are embedded within the deep learning framework to ensure physical consistency and geological interpretability.
Application results demonstrate that the proposed method achieves superior prediction performance in low–organic-matter marine shales, yielding a root mean square error of 0.08% and a coefficient of determination (R²) of 0.95. The model consistently outperforms multivariate regression, uranium-based approaches, and porosity-difference methods, while maintaining stable predictive capability in intervals exhibiting pronounced TOC heterogeneity. These results indicate that physics-constrained deep learning integrated with multi-source logging data provides a reliable and effective approach for micro-scale TOC evaluation and favorable reservoir identification in low–organic-matter marine shale systems.
This research was supported by the Natural Science Foundation of Shandong Province of China (ZR2023YQ034) and Shida Jingwei Industry-Education Integration Research Institute Project (15572259-25-ZC0607-0011).
Fig. 1. Comparison of total organic carbon (TOC) prediction results for marine shale in Well JY6, X Block, obtained from different methods.
How to cite: Ge, X., Wang, Z., He, C., Ge, X., Fan, J., Wu, X., Yang, D., and Zhai, C.: A DNN–LSTM–GPC framework for TOC prediction in marine shales of the X Block using multi-source logging data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8697, https://doi.org/10.5194/egusphere-egu26-8697, 2026.
